hadoop-common-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Runping Qi (JIRA)" <j...@apache.org>
Subject [jira] Created: (HADOOP-2032) distcp split generation does not work correctly
Date Thu, 11 Oct 2007 13:20:51 GMT
distcp split generation does not work correctly

                 Key: HADOOP-2032
                 URL: https://issues.apache.org/jira/browse/HADOOP-2032
             Project: Hadoop
          Issue Type: Bug
          Components: util
            Reporter: Runping Qi

With the current implementation, distcp will always assign multiple files to one mapper to
copy, no matter how large 
are the files. This is because the CopyFiles class uses a sequencefile to store the list of
files to be copied, 
one record per file. CopyFile class correctly generates one split per record in the sequence
file. However, 
due to  the way the sequence file record reader works, the minimum unit for splits is the
segments between the 
"syncmarks" in the sequence file. 
This results in the strange behavior that some mappers get zero records (zero files to copy)
even though their 
split lengths are non-zero, while other mappers get multiple records (multiple filesto copy)
from their split (and beyond
to the next sync mark). 

When CopyFile class creates the sequencefile, it does try to place a sync mark between splitable
segments in the sequence file by calling sync() function of the sequence file record writer.

Unfortunately, the sync() function is a no-op for files that are not block compressed.

Naturally, after I changed the compression type for the sequence file to block compression,
mappers got the correct records from their splits.
So a simple fix is to change the compression tye to CompressionType.BLOCK:

// create src list
    SequenceFile.Writer writer = SequenceFile.createWriter(
        jobDirectory.getFileSystem(jobConf), jobConf, srcfilelist,
        LongWritable.class, FilePair.class,

This message is automatically generated by JIRA.
You can reply to this email to add a comment to the issue online.

View raw message